Investigating the Impact of Rational Dilated Wavelet Transform on Motor Imagery EEG Decoding With Deep Learning Models

Decoding neural signals from electroencephalogram (EEG) data is a challenging task due to the signals' intrinsic non-stationarity, low signal-to-noise ratio, and complex spatio-temporal dynamics. The present study investigates the impact of the Rational Dilated Wavelet Transform (RDWT) for implementing a de-noising operation before applying deep learning classifiers devised to recognize Motor Imagery (MI) EEG signals. A systematic paired evaluation (with/without RDWT) is conducted on four state-of-the-art deep learning architectures and carried out across three benchmark datasets. The performance of the models integrating RDWT is reported with subject-wise averages using accuracy and Cohen's kappa, complemented by subject-level analyses to identify when RDWT is beneficial. Our results show improvements in classifier performance in almost all cases, and especially for MI tasks/trials that are difficult to classify. We then conclude that RDWT-based signal preprocessing can mitigate localized noise and enhance the EEG classification performance without adding significant complexity to the classifier architecture.